Interactive Learning Course

LLM Fundamentals

Master how large language models work -- from neurons to transformers to RLHF. Interactive exercises, code sandboxes, and AI-powered evaluation.

22

Modules Ready

25

Total Modules

0

Concepts Learned

--

Avg Score

Overall Progress: 0%

1

Architecture

9 modules -- How transformers work from the ground up

Introduction: How LLMs Work

Neural networks, forward pass, training loop, next-word prediction

Characters to tokens, BPE algorithm, byte-level encoding

The Embedding Layer

One-hot vs dense embeddings, semantic space, king-queen arithmetic

Query/Key/Value, dot products, softmax, multi-head attention

Positional Encoding

Sinusoidal encoding, RoPE, why word order matters

Layers of Understanding

Residual connections, layer normalization, FFN, depth

Learning to Predict

Cross-entropy loss, backpropagation, batches, learning rate

Instruction Tuning & RLHF

From base model to assistant, reward models, PPO

GPT-2 from Scratch

Build and train a working GPT-2 in PyTorch

2

Pre-training

8 modules -- Training at scale: data, compute, and optimization

Word2Vec to GPT-3, the evolution of pre-training

Training Objectives & Architecture

CLM vs MLM, RoPE, RMSNorm, SwiGLU, FlashAttention

Scaling Laws & Optimization

Chinchilla, AdamW, learning rate schedules, gradient clipping

Training Data Engineering

CommonCrawl, deduplication, data mixing, quality filtering

Training Infrastructure

GPU clusters, distributed training, parallelism strategies

Advanced Objectives

Curriculum learning, synthetic data, multi-stage training

Evaluation During Pre-training

Benchmarks, validation loss, emergent capabilities

Case Study: LLaMA 3

How Meta built LLaMA 3 end-to-end

3

Post-training

5 modules -- From base model to helpful assistant

Why base models aren't enough, the alignment problem

Supervised Fine-Tuning

Instruction tuning, data quality, formatting

Preference Optimization

RLHF, DPO, reward models, PPO

Tools & Safety Tuning

Tool use, function calling, safety guardrails

Case Study: Tulu 3

Open post-training recipe end-to-end

4

Advanced Topics

3 modules -- Efficiency and specialization

1

Distillation

Compressing large models into smaller, faster ones

Coming Soon

2

LoRA

Low-rank adaptation for efficient fine-tuning

Coming Soon

3

Mixture of Experts

Sparse models, routing, conditional computation

Coming Soon